Underwater Image Quality Improvement Method Based On Convolutional Neural Network

Due to the severe imaging environment and lighting conditions,underwater images contain various degradations such as low contrast,color distortion,texture blur,and non-uniform illumination.These degradations seriously affect underwater applications,e.g.,mineral exploration,underwater target detection and classification,water navigation,marine geography engineering survey and marine military.Thus,in-depth study of underwater image quality improvement is urgently required.At present,deep learning based methods have achieved excellent performance on different image processing tasks,such as rain removal,low light enhancement and defogging,at the cost of large amount of labeled data.However,it is unpractical to collect large numbers of underwater image with corresponding ground truth,which makes it hard to train a deep neural network for underwater image enhancement.To address the above issues,this work proposes a underwater image quality improvement method based on convolutional neural network.The main content and innovations are as follows:(1)To address the problem of lacking underwater images,a simple and effective style transfer algorithm is proposed to generate realistic underwater images with corresponding ground truth.The proposed algorithm is designed based on image feature matching.(2)By combining convolutional neural network and domain knowledge,a novel network architecture is designed for improving underwater image quality.The proposed network directly learns the nonlinear mapping relationship between underwater images and ground truths from the synthetic underwater image dataset.Specifically,the input image is firstly decomposed into low-frequency and high-frequency components by using a low-pass filter.Then each component is processed by individual network.The network for low-frequency parts is designed to handle global color distortion,while the one for high-frequency parts is designed to enhance local details.Experiments show that the proposed algorithm can achieve color correction,image detail reveal and sharpness in both synthetic and real-world underwater images.Moreover,the proposed algorithm has a good trade-off between effectiveness and efficiency.Since the whole operation is feed-forward after network training and can be accelerated by GPU,it has potential values for practical applications.